rho.agent

rho.agent is Rho’s provider-neutral, multi-turn agent loop. It turns provider streams into an ordered lifecycle, schedules typed tools, drains steering and follow-up queues, and propagates cancellation without placing coding or bioinformatics policy in the core.

Run an agent

library(rho.async)
library(rho.ai)
library(rho.agent)

agent <- rho_agent(
  rho_faux_provider(),
  rho_model(provider = "faux", id = "faux")
)
task <- rho_prompt(agent, "show the protocol")
rho_is_task(task)
#> [1] TRUE

run <- rho_await(task, timeout = 5000)
list(
  status = run@status,
  answer = run@messages[[2L]]@content[[1L]]@text,
  events = length(run@events)
)
#> $status
#> [1] "completed"
#>
#> $answer
#> [1] "faux: show the protocol"
#>
#> $events
#> [1] 12

Listeners are awaited in registration order. Tool calls that declare ToolMayOverlap may run concurrently, while results remain in source order; ToolRequiresExclusiveExecution makes stateful semantics explicit. Agent policy is an S7 protocol, so applications can override context transforms and before/after-tool decisions without replacing the loop.

The session ledger is append-only. rho_compact() adds a semantic checkpoint while retaining the complete transcript, and rho_build_agent_context() projects the messages sent to a provider. Threshold compaction and one retry for a typed provider input-limit value use the same compactor and policy generics.

Extensions build on this lifecycle in rho.ext. See the rho.agent reference.